TS SDK: declare task dependencies by calling tasks

[jason810496]

• next: TS SDK: generate Dag and task options from the serialization schema #73436
• Part of the native TypeScript Dag stack, TS SDK: declare task dependencies by calling tasks #73435 to TS SDK: cover native TypeScript Dags end to end #73446, to be merged in order.

Why

ts-sdk/adr/0002-native-dag-interface.md decides that the call graph is the task graph: dag.task(...) hands back a factory, and calling it places the task and supplies its arguments. Before this, wiring travelled in an inputs option nothing read.

How

• dag.task(taskId, handler) returns a TaskFactory. Calling it records the task's place and its inputs; each argument takes an upstream reference or a literal JSON value.

const extract = dag.task("extract", async (): Promise<number> => 42);
const transform = dag.task("transform", async (rows: number, region: string) => rows);
const load = dag.task("load", async (total: number) => {});

load(transform(extract(), "us"));

• Arguments are passed in the order the handler declares them. A handler that declares a single object of named arguments can also be called with that object, which is the shape Python TaskFlow uses for load(transformed=transform(...)).
• A reference passed by position is checked against the argument's own type, which is what tells the two call shapes apart when a handler declares a single argument.
• A reference has to be the argument itself. One buried in an array or an object draws no edge, so it is rejected rather than recorded as data.
• Every task has to be called exactly once, so none is silently left out of the graph. A TaskRef exists only once its producing call has returned, which is what makes a cycle through arguments unwritable.
• A TaskRef carries its handler's return type, so a construct that needs a particular one can ask for it.
• Bundle keeps the Dags declared in TypeScript and the task handlers for Python-declared Dags in separate maps, since one dag_id belongs to one of them and never both.

---

Was generative AI tooling used to co-author this PR?

• Yes, with help of Claude Code Opus 5 following the guidelines

[pierrejeambrun]

Overall LGMT.

Just one nit.